Open Advanced Search
Get 20M+ Full-Text Papers For Less Than $1.50/day.
Start a 14-Day Trial for You or Your Team.
Learn More →
A Before-and-After Study of a Collision Risk Detecting and Warning System on Local Roads
A Before-and-After Study of a Collision Risk Detecting and Warning System on Local Roads
Yun, Jeongin;Selmoune, Aya;Chae, Junseong;Seo, Myoungkook;Lee, Jinwoo
Hindawi Journal of Advanced Transportation Volume 2023, Article ID 6894065, 13 pages https://doi.org/10.1155/2023/6894065 Research Article A Before-and-After Study of a Collision Risk Detecting and Warning System on Local Roads 1 2 3 4 Jeongin Yun , Aya Selmoune , Junseong Chae , Myoungkook Seo , and Jinwoo Lee Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology, Daejeon 34051, Republic of Korea Jiangsu Key Laboratory of Urban ITS, Southeast University, Nanjing 210000, China Research and Development Center, Pintel Incorporated, Seoul 06729, Republic of Korea Smart Engineering Laboratory, Korea Construction Equipment Technology Institute, Gunsan 54004, Republic of Korea Correspondence should be addressed to Jinwoo Lee; firstname.lastname@example.org Received 17 September 2022; Revised 15 January 2023; Accepted 20 January 2023; Published 28 February 2023 Academic Editor: Jaeyoung Lee Copyright © 2023 Jeongin Yun et al. Tis is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Local roads have numerous blind spots caused by complex geometry, obstacles, and narrow width. Tus, conventional proactive countermeasures, such as passive trafc signs and convex mirrors, have not always been efective in preventing local road collisions. In this paper, we present a novel proactive two-step approach for trafc safety on local roads, comprised of detection of pedestrian-to-vehicle and vehicle-to-vehicle collision risks and warning systems. First, using video surveillance and radars to eliminate blind spots, the system detects road objects, predicts their trajectories and reachable areas, and identifes a potential risk situation. Second, it provides road users such as vehicles and pedestrians with warnings through LED variable message signs, which allows them to react efectively in risky situations. We have applied the system to two local road sites in South Korea, including a university campus in Seoul City and an apartment complex in Daejeon City. Te detecting system has been validated using a confusion matrix. We have assessed the warning efect through a before-and-after study and found that the proposed system contributed to the improvement of trafc safety at the case study site in that trafc conficts decreased by 55–62%. essential. Recently, pedestrian safety has emerged as a global 1. Introduction public health priority, as pedestrians are exposed to a variety According to the global status report on road safety pub- of threats and are prone to sustaining severe injuries in trafc lished by World Health Organization, the annual number of accidents. Since pedestrian trafc is concentrated on local road trafc fatalities reached 1.35 million in 2016, and there roads, pedestrian safety control on local roads is particularly have been approximately 1.3 million every year since [1, 2]. important. However, due to numerous blind spots on local roads caused by complex geometry, obstacles, and narrow More than 50% of all road trafc fatalities occur among vulnerable road users, including pedestrians, bikers, and width, it is difcult for road users to accurately estimate their motorcyclists. Annually, between 20 and 50 million people current safety situations. A blind spot is defned as the spatial have nonfatal injuries, with many developing disabilities as a area outside of a person’s peripheral vision, and it restricts result. Individuals, their families, and nations as a whole drivers and passengers from seeing objects and other pe- incur signifcant economic losses due to road trafc acci- destrians moving. Tis is one of the main reasons that dents. Te average cost of road trafc accidents is 3% of a traditional proactive countermeasures, such as trafc signs country’s gross domestic product. and convex mirrors, have not always been efective to For the efcient and efective implementation of trafc prevent local road collisions. To overcome this issue, an safety measures, adequate trafc safety management is advanced proactive countermeasure is required that can 2 Journal of Advanced Transportation information, which allows them to react efectively in risky eliminate blind spots and provide accurate safety infor- mation in advance to local road users, including pedestrians. situations. We have validated the complete form of a safety countermeasure, including detection and warning systems Proactive near miss accident warning technologies have received much attention in the past decade, and numerous through implementing it on real-world local road sites in sensor-based algorithms have been developed for a better South Korea, including a university campus in Seoul City and understanding of current safety situations. Tese algorithms an apartment complex in Daejeon City. We have validated the recognize objects, estimate their trajectories, and assess system using a confusion matrix and assessed the trafc safety whether a collision risk exists; based on surrogate safety efect through a before-and-after (BA) study. measures (SSMs) derived from trafc conficts, they estimate Te remainder of this paper is structured as follows: the system description section provides an overview of the the collision risk of particular trafc scenarios using mi- croscopic trafc parameters such as vehicle speed, acceler- proposed system’s architecture. Next, we elaborate on the case study sites. Te evaluation results section describes the ation, time headway, and space headway. Depending on how the collision risk is evaluated, the algorithms can be cate- process and outcomes of the analysis. In the last section, we present the conclusion with future research directions. gorized into two groups. Te frst group of algorithms uses deep learning methods to detect hazardous situations; after training with prior data 2. System Description categorized as risk situations by SSMs, algorithms predict whether a current circumstance is risky or not. For instance, In this section, we describe how the proposed system works. gated recurrent unit and long-short term memory (LSTM) Te system consists of four steps: (1) object detection and approaches were used to forecast the likelihood of a collision perspective transformation; (2) trajectory and reachable area at a signalized intersection [3, 4]. For unsignalized cross- prediction; (3) collision risk determination; and (4) collision walks, another LSTM-based collision risk area estimation risk warning. method was proposed . Te second group of algorithms applies SSMs to notice the presence of potentially hazardous scenarios based on the 2.1. Object Detection and Perspective Transformation. In this overlap of expected object trajectories. Recent years have step, object detection and perspective transformation are witnessed a surge of scholarly works examining SSM practices conducted. First, for object detection, we use a video sur- and/or trafc conficts, such as [6–8]. SSMs, including time- veillance system and radar devices. We develop the object- to-collision (TTC) and post encroachment time, have been detecting algorithm based on You Only Look Once (YOLO) extensively utilized to assess trafc safety performance and v5, which is the latest deep-learning method and has been identify potential accident hazards. For example, Son et al. used for real-time video-based detection [16, 17]. Te de- developed an algorithm to detect the risk of collisions between veloped detecting algorithm is diferent from the YOLO trucks and pedestrians . Using microscopic simulations, algorithm in terms of how data is processed and analyzed. the authors established a road network, and the algorithm was Te size of input/output image data is not fxed but rather then validated using a confusion matrix. Moreover, Zhang optimized based on hardware usage. In addition, input data et al. proposed a novel multipedestrian collision risk as- are not processed in random order but in a frst-in-frst-out sessment approach that consists of collision checking, motion order so that image analysis proceeds without a bottleneck, prediction, and collision risk assessment modules, and Wu as illustrated in Figure 1. Moreover, the data to be analyzed et al. developed a crash warning system for bicycle lane areas are processed as metadata that is compressed image infor- at intersections utilizing connected vehicle technology mation. It can improve detection accuracy and computing [10, 11]. Ke et al. presented an approach to automatically speed without image quality and pixel loss. detect near collisions between vehicles and pedestrians using Our proposed object-detecting algorithm was com- onboard monocular vision and a moving background . pared to the diferent versions of YOLO, including According to the extant literature, collision detection and/or YOLOv4 TRT, YOLOv4 Darknet, and YOLOv5m. Tey prevention systems are mostly vehicle-integrated; recent are considered to be superior to other traditional studies have reviewed the threat assessment methods for such detecting methods, such as single-shot detection and systems [13–15]. convolutional neural network algorithms, in terms of high However, to the best of our knowledge, few algorithms accuracy and computational speed [18–20]. Figure 2 il- have been developed specifcally for local roads with blind lustrates object-detecting performance for the existing spots. Moreover, no study has implemented and assessed the YOLO algorithms and the proposed method with respect safety impacts of their proposed algorithms in real-world to the mean average precision (mAP) and frames per practice. In this paper, we present a novel proactive two-step second (FPS), commonly used to evaluate the accuracy approach for trafc safety on local roads, comprised of de- and computing speed of detecting methods [21, 22]. Te tection of pedestrian-to-vehicle and vehicle-to-vehicle colli- proposed detection algorithm has higher accuracy than sion risks and warning systems. We implemented video and YOLOv5m, but their computing speeds are comparable. radar equipment to eliminate blind spots at a target local road In terms of both mAP and FPS, the proposed algorithm site. In addition to the detection system, we implemented outperforms the rest detection algorithms. Consequently, LED Variable Message Signs (LED-VMS) for providing road based on the aforementioned comparison results, it can be users such as vehicles and pedestrians with warning concluded that the proposed system is superior to Journal of Advanced Transportation 3 Random order Detector 1 First-in-first-out Batch 1 Detector 1 order Camera 1 Batch 1 Camera 1 Detector 2 GPU Batch 2 2 Detector 2 GPU Camera 2 Batch 2 Camera 2 Data queue Detector m Batch m Detector m Camera m Batch m Camera m <Analytic server> <Analytic server> (a) (b) Figure 1: Data processing method: (a) random order process method and (b) frst-in-frst-out order process method (proposed system). 17.9 97.0 18.0 17.1 93.2 83.4 74.2 6.1 70.6 63.3 62.9 Small Medium Large Proposed YOLOv4 YOLOv4 YOLOv5m model Darknet TRT Object size Proposed model Proposed model YOLOv4 TRT YOLOv4 Darknet YOLOv4 Darknet YOLOv5m YOLOv4 TRT YOLOv5m (a) (b) Figure 2: Object detecting performance comparison: (a) mean average precision (mAP) and (b) frames per second (FPS). traditional detecting methods. Moreover, radar devices time points, respectively. Te degree of curvature θ [ra- are employed when video-based detection is unavailable, dians] at the time frame n is calculated using the following such as in extreme weather conditions. From the video equation: and radar equipment, the location, speed, and type of the θ � θ · ω , n n−i i (1) object can be precisely calculated. Second, a perspective i�1,... transformation is performed. We calculate the objects’ overhead coordinates from the raw data collected from the where i is the backward of the time frame index and ω is a sensors’ perspective. Terefore, we use a perspective predetermined weight factor such that ω > ω ′ for all i < i . i i transformation matrix in the Open CV library, which is a If |θ | < θ , an object will move straight. Te object’s 0 set computer vision tool . Te transformation procedure future location is determined based on the expected moving is depicted in Figure 3. distance calculated by the following equation: d � t − t v n n 0 0 (2) future 2.2. Trajectory and Reachable Area Prediction. Tis step is to ∀n ∈ 1, . . . , T , predict the future locations of the detected objects. First, we identify whether an objective’s prior trajectory is straight or where d is the expected moving distance [meters] between the curved, depending on the degree of curvature of the current object’s center positions at the current and future time points t trajectory. If it is larger than a predetermined threshold θ and t , respectively; v is the estimated current velocity; and set n 0 future [radians], we consider the trajectory to be curved. Te time T is the prediction time frame horizon. If |θ | ≥ θ , the 0 set future location at the time point t is determined based on the frame index is denoted k, and the current frame is k � 0. Te negative and positive k values mean the prior and future location at the time point t , v , and θ . n−1 0 n−1 mAP (mean average precision) (%) FPS (frames per second) 4 Journal of Advanced Transportation (x , y ) 1 1 (x′, y′) (x′, y′) 0 0 1 1 (x , y ) 2 2 (x , y ) 0 0 (x′, y′) (x′, y′) 2 2 3 3 (x , y ) 3 3 (a) (b) Figure 3: Perspective transformation: (a) coordinates from video surveillance systems and (b) overhead perspective coordinates. Considering the actual size and uncertain future To reduce the computational burden for calculation be- movement of objects, we set the reachable area of an object at tween multiple objects on roads, a following three-step al- k � n as a rectangle and an ellipse for vehicles and pedes- gorithm is used, as illustrated in Figure 6. In the frst step, we trians, respectively. Te vehicles’ rectangular lengths, l set the maximum detection distance for each object, as shown major [meters] and l [meters], are determined by the fol- in Figure 6(a). If a distance from the object (object A in minor Figure 6(a)) to another is greater than the maximum detection lowing equation based on a previous study : distance (e.g., thirty meters), we exclude the objects. Second, l � max0.61d − 1.06,5.0, major n as depicted in Figure 6(b), we exclude situations when the (3) future trajectories of two objects intersect outside local roads l � max0.57d + 0.96,2.5. minor n or when the time diference between the two objects’ arrival A pedestrian’s future location is stochastically defned by times to the intersecting point exceeds a predetermined an ellipse . Te ellipse major axis length, r [meters], threshold (e.g., two seconds). Last, we check if the reachable major and minor axis length, r [meters], are empirically es- areas of two objects (not their trajectories) overlap in 0.25- minor timated by pedestrian trajectory data collected from video second increments starting from one second earlier than the surveillance systems on local roads in Korea, as illustrated in earlier time between the predicted arrival times of two objects Figure 4(a) . Pedestrian trajectories were plotted every at the trajectory intersection. If there is an overlap three times second for fve seconds. It was found that the length of the or more consecutively, we consider the situation a collision major axis and minor axis varied depending on d . From risk and defne TTC as the time when the two reachable areas plotted trajectories, r was linearly modeled, as illustrated frst overlap, as shown in Figure 6(c). major in Figure 4(b). For r , trajectories were plotted mostly minor within two meters. Te estimated ellipse axis lengths are 2.4. Collision Risk Warning. If a risk situation is identifed calculated using the following equation : and its predicted TTC is shorter than a predefned threshold, such as four seconds, a warning message is displayed on the r � max1.6d − 0.89,1.0, major n LED-VMS, as depicted in Figure 7. TTC threshold is ob- tained as the total summation of the perception-reaction 2.0, d ≥ 3.0 , ⎧ ⎪ (4) time , a safety margin time, and the time for the vehicle r � stopping distance. As shown in Figure 7, a warning message minor indicating “Stop” is provided. Te LED-VMS is normally of, d + 1.0, d < 3.0 . n n but when a risk situation occurs, it provides warning in- formation until the risk situation is over. Te LED-VMS is installed mainly in locations where they can be seen directly 2.3. Collision Risk Determination. If a risk is defned based in the direction of vehicle and pedestrian routes, allowing drivers and pedestrians to get warning information easily. solely on proximity, we do not account for other factors possibly important for risk identifcation, such as speed, di- rection, and future location uncertainty. Tus, we defne the 3. Case Study Sites collision risk between two objects when their reachable areas overlap, which are obtained based on not only proximity but We implemented the proposed system to two local road also the other factors since it has been proven to reduce false networks, a university campus in Seoul City and an apartment negative or positive cases . As illustrated in Figure 5, the complex in Daejeon City, South Korea. First, we selected one proposed system detects pedestrian-to-vehicle and vehicle-to- site on the university campus for system verifcation, as shown vehicle collision risks. Pedestrian-to-pedestrian collision risks in Figure 8, and we validated the system using a confusion are excluded since it is associated with relatively low severity, matrix before the evaluation. Second, we selected two pairs of and pedestrians can evade a pedestrian-to-pedestrian collision experimental and control sites in the apartment complex in promptly. Tus, the system does not consider the overlap of Daejeon City for a BA study. Te sites of each pair are as- pedestrians’ reachable areas. sociated with similar environmental and trafc conditions Journal of Advanced Transportation 5 4 8 minor major 0 2468 10 -1 -2 -3 1 23456 -4 x [meters] d [meters] x: Pedestrian movement direction y: Vertical direction of the pedestrian movement direction (a) (b) Figure 4: Pedestrian future reachable area: (a) estimated ellipse major and minor axes from plotted pedestrian trajectory data and (b) modeled r according to d . major n Vehicle reachable area Pedestrian reachable area Overlap of predicted reachable area (a) (b) Figure 5: Collision risk defnition of the proposed system. Te system detects collision risks of pedestrian-to-vehicle and vehicle-to-vehicle as in (a), whereas the system does not account for pedestrian-to-pedestrian collision risks as in (b). [27, 28]. Te proposed system was implemented and operated and not risky, respectively. False positive (FP) indicates that at experimental sites but not at control sites. Since the ex- the actual situation is not risky, whereas the predicted perimental and control sites are located within the same situation is. On the other hand, false negative (FN) indi- apartment complex, population characteristics and temporal cates the opposite situation to FP. Using the four measures, trafc patterns on local roads can be similar between the ex- we calculate three metrics that indicate the system’s per- perimental and control sites. Te experimental and control sites formance using equations (5)–(7). Accuracy is the most of pair A have parking lots and underground parking en- intuitive measure, indicating how precisely a system trances, and the experimental and control sites of pair B have classifes situations. Te true positive rate (TPR), com- roadside parking lots, as shown in Figure 9. monly known as recall (or sensitivity), is the proportion of real-risk situations correctly predicted. It indicates how 4. Evaluation Results accurately the system classifes risky situations. On the other hand, the false positive rate (FPR) indicates incor- 4.1. Confusion Matrix. We use a confusion matrix approach rectly predicted no-risk situations among actual no-risk to validate the performance of the implemented system, as situations. If the FPR is high, it indicates that false alarms shown in Table 1. True positive (TP) and true negative (TN) can occur despite the absence of potential risk situations. indicate that both predicted and actual situations are risky TP + TN Accuracy � , (5) TP + TN + FP + FN TP (6) True Positive Rate (TPR) � Recall (or Sensitivity) � , TP + FN y [meters] r [meters] major 6 Journal of Advanced Transportation 30 m Detected objects (Within 30 meters of A) (a) No collision risk 0.5 seconds 3 seconds Collision risk No collision risk (b) Predicted collision risk t seconds t < t n m t seconds t – 1 (t – 1) + 0.25 (t – 1) + 0.50 (t – 1) + 0.75 n n n n <Present = t > <After t – 1 seconds> 0 n It is determined as a collision risk (TTC is defined as (t – 1) + 0.25 seconds) (c) Figure 6: Tree-step method to identify risky situations. (a) Two objects are excluded as the distances from object A exceed thirty meters. (b) Intersecting points (blue dots) are excluded since the point is outside local roads or the time diference in arrival times to the point exceeds two seconds. (c) If reachable areas overlap three times consecutively, the situation is considered to be a collision risk, and TTC is defned as the time when the two reachable areas frst overlap. Journal of Advanced Transportation 7 A collision risk happens No risk situation Risk situation A collision risk ends (a) (b) Figure 7: LED-VMS provides warning information about collision risks. (a) In a normal situation, LED-VMS is of. (b) In a risk situation, LED-VMS is turned on until the risk situation is over. Site for validation (a) (b) Figure 8: System application (Seoul campus): (a) map of the site for validation and (b) image recorded by a video surveillance camera on Seoul campus. FP (7) False Positive Rate (FPR) � 1 − Specificity � . FP + TN We collected data on risk and normal situations from the from August 1, 2022, to August 2, 2022. Te sample numbers sensors implemented on the university campus in Seoul including both risk and no-risk situations were 162 and 135 City. We have tested the model’s performance under two for the clear and rainy weather periods, respectively, and the diferent weather conditions, clear and rainy, and two sample numbers for day and night conditions are 235 and lighting conditions, day and night. For example, photos of 62, respectively. Table 2 presents the system verifcation results for weather and lighting conditions. Under clear one application site under clear and rainy weather are shown in Figure 10. Te data collection period for clear weather was weather, accuracy, TPR, and FPR were found to be 80.9%, from July 8, 2022, to July 11, 2022, for forty-eight hours. Te 76.1%, and 15.4%, whereas, under rainy weather, they were same amount of time data was collected on rainy weather 80.0%, 77.4%, and 18.3%, respectively. 8 Journal of Advanced Transportation Experimental Control site A site A Control site B Experimental site B (a) (b) (c) (d) (e) Figure 9: System application at the apartment complex in Daejeon City. (a) Map of two site pairs; (b, c) parking lot and underground parking entrance (experimental and control sites A); (d, e) roadside parking lot (experimental and control sites B). Te fndings confrmed that the performance of the performance [3, 4, 9]. However, there was a diference be- system particularly for local roads is comparable to that of tween day and night. In the nighttime, accuracy and TPR were other systems for other kinds of roads in terms of overall lower than in the daytime, while FPR was higher than in the Journal of Advanced Transportation 9 Table 1: Confusion matrix. Actual situation Classifcation Risk No risk Risk True positive (TP) False positive (FP) Predicted situation No risk False negative (FN) True negative (TN) (a) (b) Figure 10: Application site (a campus in Seoul city) under diferent weather conditions: (a) clear weather and (b) rainy weather. Table 2: System verifcation results under diferent weather and lightning conditions. Site Weather (number of samples) Accuracy (%) TPR (%) FPR (%) Clear (162) 80.9 76.1 15.4 Rainy (135) 80.0 77.4 18.3 Day (235) 82.6 78.2 14.2 Night (62) 72.6 69.6 25.6 University campus in Seoul city Clear-day (133) 82.7 77.6 13.3 Clear-night (29) 72.4 69.2 25.0 Rainy-day (102) 82.4 79.1 15.3 Rainy-night (33) 72.7 70.0 26.1 Total (297) 80.4 76.6 16.8 daytime. It implies that the lighting condition can afect the certain technical approach. Typical BA studies exploit his- torical collision data to evaluate the changes in collision system’s performance , and improving the performance even at nighttime remains a future work. For the comparison frequency and/or severity attributable to safety interventions of clear and rainy weather, there was no discernible diference . Several collision-based BA studies have adopted the between sunny and wet conditions in terms of either accuracy empirical and the full Bayes methods [32–35]. or TPR. However, the FPR was marginally higher under wet Nonetheless, due to the rarity and randomness of crashes, conditions than in dry conditions. Tis implies that there is a statistically reliable safety evaluations require extended pe- greater likelihood of false alarms occurring while it is raining, riods of collision data before and after the implementation of since wet circumstances can limit vision and image clarity, safety measures, particularly in the case of local road colli- hence lowering the precision of object recognition and tra- sions. Terefore, the use of collisions in BA studies may raise a jectory prediction . In addition, small errors measured by moral confict of waiting for “collisions” to occur before the sensors (video surveillance camera and radar) can attempting to prevent them. Moreover, the use of crash data has usually been limited by sample size and under-reporting compromise the object detection and trajectory prediction process. Tus, the model could erroneously categorize a no- . For that reason, using crash data is considered a reactive risk situation as risky. analysis method as opposed to a proactive one. Hence, in this study, we used trafc conficts to conduct a BA study. Trafc confict-based analysis provides insight 4.2. Before-and-After Study. Before-and-after safety inves- into the failure mechanism of trafc accidents with a marginal social cost . Additionally, it does not require a tigation is an essential element of road safety improvement schemes that strive to quantify the potential benefts of a long observation period. In July of 2022, we collected trafc 10 Journal of Advanced Transportation Experimental site A Control site A 70 70 60 60 50 50 Before Before After After (a) Experimental site B Control site B 120 120 80 80 60 60 40 40 20 20 Before Before After After (b) Figure 11: Number of trafc conficts in before-and-after periods: (a) experimental and control sites A (parking lot and underground parking entrance) and (b) experimental and control sites B (roadside parking lot). Before After confict data from the surveillance systems installed at two experimental site of pair j and C and C are those j j pairs of experimental and control sites; the sample sizes for numbers for before and after the experimental periods at the pairs A and B were 253 and 597, respectively, which are control site of pair j. In this study, before-and-after ex- statistically reliable based on the method proposed by periments correspond to whether or not LED-VMS pro- [38, 39]. Te observed trafc confict frequencies at indi- vided road users with warning information. A value of OR vidual sites are shown in Figure 11. Te fgure shows that, smaller than one shows that the experiment is efective while for all experimental sites, the confict frequency was sub- OR values greater than one show a negative efect. When OR stantially reduced, whereas, for all control sites, the number equaled one, all changes are attributable to factors unrelated of trafc conficts did not difer in the before-and-after to the experiment. To intuitively examine the efect, we period. calculated the treatment efect (TE) for a pair j using the To statistically analyze the efect, we adopted the odds following equation: ratio (OR) method [27, 40–42]. Tis method is usually TE � OR − 1, (9) j j employed to systematically compare the efects on experi- mental sites and control sites. By using the OR method, we where TE represents the change in the frequency of trafc can explain before-and-after temporal changes in trafc conficts at the experimental site of pair j after the imple- conficts and exclude factors unrelated to safety interven- mentation of the proposed system, considering the change in tions . Te OR indicator for pair j, OR , can be cal- the frequency of trafc conficts at the control site of pair j. We can calculate the percentage of reduction in trafc culated as in the following equation: conficts as 100(%) × TE. Negative values of TE mean a After Before E /E j j reduction in conficts—in other words, a safety improve- OR � , (8) After Before ment. On the contrary, positive values of TE indicate an C /C j j increase in conficts and, thus, a safety deterioration. Table 3 Before After where E and E represent the number of trafc shows the values of OR and TE. Both A and B experimental j j conficts in before-and-after experimental periods at the sites showed negative TE. Journal of Advanced Transportation 11 Table 3: Odds ratio and treatment efect at site pairs A and B. Site Period Confict frequency Odds ratio (OR) Treatment efect (TE) Before 61 Experimental site A After 24 0.376 −0.624 Before 64 Control site A After 67 Before 149 Experimental site B After 64 0.450 −0.550 Before 154 Control site B After 147 We integrated the individual ORs using a weighted Table 4: Trafc confict reduction (safety efect). average to estimate the total OR. In addition, we assessed the Experimental site Confict reduction (%) p value statistical signifcance of the calculated ORs by testing the A 62.4 0.0005 following hypothesis: OR is equal to one (null hypothesis). − 5 B 55.0 1.168 × 10 We proceeded by converting OR to a logarithmic form. OR − 8 Total 57.3 4.852 × 10 is always positive because the number of trafc conficts cannot be negative. Tis observation led to the generaliza- tion that OR should follow a lognormal distribution. A suggests that the efect is greater at experimental site A than standard error (SE) of ln(OR ) for an individual site pair can at experimental site B. Tis is presumably due to the location be estimated as in the following equation: of LED-VMS at both sites. At experimental site A, the LED- ������������������������� VMS is located in front of the underground parking exit. 1 1 1 1 However, at experimental site B, the LED-VMS is located on SE � + + + , (10) After Before After Before E E C C the roadside. Vehicles exiting the underground parking lot j j j j in experimental site A can see the LED-VMS signal more where SE is a standard error in site pair j. In addition, we clearly than vehicles in experimental site B, which can have assumed that weights are inversely proportional to the an impact on driver reaction time. variance of individual OR. A weight factor to calculate total OR can be calculated as in the following equation: 5. Conclusions − 1 1 1 1 1 1 ⎛ ⎝ ⎞ ⎠ In this study, we present a novel proactive two-step ap- w � � + + + , (11) 2 After Before After Before E E C C SE proach for trafc safety on local roads, comprised of de- j j j j tection and warning. First, using video surveillance and where w is a weight factor in the site pair j. Te total OR for j radars to eliminate blind spots, the system detects objects two site pairs can be calculated as in the following equation: and predicts their trajectories and reachable areas; the system then uses time-to-collision to identify potential risk w lnOR j j j�1 scenarios based on the overlap of predicted reachable areas. ln(OR) � . (12) w Second, the system provides LED-VMS-based warnings that j�1 enable road users, such as drivers and pedestrians, to re- Te test statistic z asymptotically follows a standard spond efectively in specifc circumstances. normal distribution. It can be calculated as in the following We installed and operated the proposed system on a equation: university campus and an apartment complex in South ������ � Korea. First, we validated the system’s performance using a z � ln(OR) w ∼ N(0, 1). (13) j�1 confusion matrix; the system was accurate 80% of the time, and 77% of collision risks were successfully detected by the We reject the null hypothesis if the tail probability of the system, whereas 17% were false alarms. Second, we assessed probability density function is smaller than the signifcance the trafc safety efect through a BA study. We chose two level of 0.05. Te statistical signifcance result of the safety pairs of experimental and control sites with trafc and efect is presented in Table 4. Experimental sites A and B geometry comparable. Te proposed system signifcantly showed 62.4% and 55.0% reductions in trafc conficts, reduced trafc conficts and improved trafc safety, as the considering control sites A and B. Te number of conficts results indicated a decrease in trafc conficts ranging from decreased by 57.3% overall. All of the results were found to 55% to 62% across various locations. be statistically signifcant. It is, therefore, implied that the Te fndings from this research are encouraging, and proposed system improved trafc safety at all experimental further work needs to be carried out to improve safety on local sites. roads. We are aware that our research may have some limi- In experimental site A, confict reduction was approx- tations. Local roads are difcult to standardize like highways or imately 62.4%, which is greater than that in experimental site higher road hierarchies than local roads, and implementing the B, where confict reduction was approximately 55.0%. It proposed system on various local road sites requires a high 12 Journal of Advanced Transportation based on long short-term memory neural network,” Accident amount of budget. Tus, we do not intend to estimate standard Analysis & Prevention, vol. 148, Article ID 105799, 2020. crash modifcation factors for the proposed system on local  B. Noh and H. Yeo, “A novel method of predictive collision roads through a BA study. Instead, we focus on two sites in the risk area estimation for proactive pedestrian accident pre- context of a case study to derive practical insights from sta- vention system in urban surveillance infrastructure,” Trans- tistically proven site-specifc data analysis results. Currently, portation Research Part C: Emerging Technologies, vol. 137, this research consortium is planning to extend the number of Article ID 103570, 2022. testing sites in the next few years to improve the generality of  C. Johnsson, A. Laureshyn, and T. De Ceunynck, “In search of the proposed system’s potential benefts. In the future, we will, surrogate safety indicators for vulnerable road users: a review therefore, be able to collect and examine a higher number of of surrogate safety indicators,” Transport Reviews, vol. 38, samples from newly exploited sites to conduct a meta-analysis no. 6, pp. 765–785, 2018. for an objective measurement of the integrated quantitative  A. Arun, M. M. Haque, A. Bhaskar, S. Washington, and evidence. Second, warning information from LED-VMS alone T. Sayed, “A systematic mapping review of surrogate safety as- sessment using trafc confict techniques,” Accident Analysis may not be sufcient for pedestrians using smartphones while & Prevention, vol. 153, Article ID 106016, 2021. walking. Audible alerts would be more suitable for such in-  C. Wang, Y. Xie, H. Huang, and P. Liu, “A review of surrogate stances. Tus, we will install multiple heterogeneous warning safety measures and their applications in connected and systems that will be used to provide road users with risk in- automated vehicles safety modeling,” Accident Analysis & formation. In addition, we are planning to further optimize the Prevention, vol. 157, Article ID 106157, 2021. accuracy and efectiveness of the system by considering confict  S. o. Son, J. Park, C. Oh, and C. Yeom, “An algorithm for severity assessment, to give diferent levels of warnings (for detecting collision risk between trucks and pedestrians in the instance, when the risk is low, only LED-VMS is used, whereas connected environment,” Journal of Advanced Trans- when the risk is high, both LED-VMS and an audible warning portation, vol. 2021, pp. 1–9, 2021. are sent). Te system can have multiple warning thresholds to  L. Zhang, K. Yuan, H. Chu et al., “Pedestrian collision risk classify the level of risk situations. assessment based on state estimation and motion prediction,” IEEE Transactions on Vehicular Technology, vol. 71, no. 1, pp. 98–111, 2022. Data Availability  Y. Wu, M. Abdel-Aty, O. Zheng, Q. Cai, and L. Yue, “De- veloping A crash warning system for the bike lane area at in- Te data used to support the fndings of this study are tersections with connected vehicle technology,” Trans- available from the corresponding author upon request. portation Research Record, vol. 2673, no. 4, pp. 47–58, 2019.  R. Ke, J. Lutin, J. Spears, and Y. Wang, “A cost-efective Conflicts of Interest framework for automated vehicle-pedestrian near-miss de- tection through onboard monocular vision,” in Proceedings of Te authors declare that they have no conficts of interest. the IEEE Conference on Computer Vision and Pattern Recog- Nition Workshops, Honolulu, United Stated, July 2017.  J. Dahl, G. R. de Campos, C. Olsson, and J. Fredriksson, Authors’ Contributions “Collision avoidance: a literature review on threat-assessment techniques,” IEEE Transactions on Intelligent Vehicles, vol. 4, Jeongin Yun and Aya Selmoune contributed equally to this no. 1, pp. 101–113, 2019. paper.  A. Mukhtar, L. Xia, and T. B. Tang, “Vehicle detection techniques for collision avoidance systems: a review,” IEEE Acknowledgments Transactions on Intelligent Transportation Systems, vol. 16, no. 5, pp. 2318–2338, 2015. Tis work was funded by the Ministry of the Interior and  M. I. Mushtaq and M. Kansal, “A review on vehicle collision Safety (MOIS), Republic of Korea (grant number 2021- avoidance system in networks,” International Research MOIS41-001-00000000-2022 (development of trafc safety Journal of Engineering and Technology, vol. 3, 2016, http:// risk warning technology for blind spots in living roads)). www.irjet.net.  F. Zhou, H. Zhao, and Z. Nie, “Safety helmet detection based on YOLOv5,” in Proceedings of the IEEE International Con- References ference on Power Electronics, Computer Applications (ICPECA), pp. 6–11, Shenyang, China, January 2021.  World Health Organization, Global Status Report on Road  P. Jiang, D. Ergu, F. Liu, Y. Cai, and B. Ma, “A Review of Yolo Safety, World Health Organization, Geneva, Switzerland, algorithm developments,” Procedia Computer Science, vol. 199, pp. 1066–1073, 2022.  World Health Organization, Road Trafc Injuries 2022, World Health Organization, Geneva, Switzerland, 2022, https://  J. A. Kim, J. Y. Sung, and S. H. Park, “Comparison of Faster- RCNN, YOLO, and SSD for real-time vehicle type recogni- www.who.int/news-room/fact-sheets/detail/road-trafc- injuries. tion,” in Proceedings of the 2020 IEEE International Confer- ence on Consumer Electronics-Asia (ICCE-Asia), pp. 1–4,  S. Zhang, M. Abdel-Aty, Y. Wu, and O. Zheng, “Modeling pedestrians’ near-accident events at signalized intersections Seoul, Korea, November 2020.  X. Zhang, W. Yang, X. Tang, and J. Liu, “A fast learning using gated recurrent unit (GRU),” Accident Analysis & Prevention, vol. 148, Article ID 105844, 2020. method for accurate and robust lane detection using two-stage feature extraction with YOLO v3,” Sensors, vol. 18, no. 12,  S. Zhang, M. Abdel-Aty, Q. Cai, P. Li, and J. Ugan, “Prediction of pedestrian-vehicle conficts at signalized intersections p. 4308, 2018. Journal of Advanced Transportation 13  K. R. Ahmed, “Smart pothole detection using deep learning  A. Høye, “Safety efects of fxed speed cameras—an empirical based on dilated convolution,” Sensors, vol. 21, no. 24, p. 8406, Bayes evaluation,” Accident Analysis & Prevention, vol. 82, pp. 263–269, 2015.  L. Zheng, T. Sayed, and F. Mannering, “Modeling trafc  R. Padilla, S. L. Netto, and E. A. Da Silva, “A survey on conficts for use in road safety analysis: a review of analytic performance metrics for object-detection algorithms,” in methods and future directions,” Analytic methods in accident Proceedings of the 2020 international conference on systems, research, vol. 29, Article ID 100142, 2021. signals and image processing (IWSSIP), pp. 237–242, Niteroi,  L. Zheng and T. Sayed, “A full Bayes approach for trafc Brazil, July 2020. confict-based before–after safety evaluation using extreme  X. Wu, D. Sahoo, and S. C. Hoi, “Recent advances in deep value theory,” Accident Analysis & Prevention, vol. 131, learning for object detection,” Neurocomputing, vol. 396, pp. 308–315, 2019. pp. 39–64, 2020.  R. B. Dell, S. Holleran, and R. Ramakrishnan, “Sample size  B. Noh, W. No, and D. Lee, “Vision-based overhead front determination,” ILAR Journal, vol. 43, no. 4, pp. 207–213, point recognition of vehicles for trafc safety analysis,” in Proceedings of the 2018 ACM International Joint Conference  A. S. Singh and M. B. Masuku, “Sampling techniques & and 2018 International Symposium on Pervasive and Ubiq- determination of sample size in applied statistics research: an uitous Computing and Wearable Computers, pp. 1096–1102, overview,” International Journal of economics, commerce and Singapore, October 2018. management, vol. 2, no. 11, pp. 1–22, 2014.  M. Seo, H. Shin, H. Jeong, and J. Chae, “Development of an  L. Zheng and T. Sayed, “Application of extreme value theory object collision detection algorithm for prevention of collision for before-after road safety analysis,” Transportation Research accidents on living roads,” Journal of Drive and Control, Record, vol. 2673, no. 4, pp. 1001–1010, 2019. vol. 19, no. 3, pp. 23–31, 2022.  J. Autey, T. Sayed, and M. H. Zaki, “Safety evaluation of right-  K. S. Oh, S. Y. Park, J. H. Seo, G. H. Lee, and K. S. Yi, “Laser- turn smart channels using automated trafc confict analysis,” Scanner-based stochastic and predictive working-risk-as- Accident Analysis & Prevention, vol. 45, pp. 120–130, 2012. sessment algo-rithm for excavators,” Journal of Drive and  P. Reyad, E. Sacchi, S. Ibrahim, and T. Sayed, “Trafc con- Control, vol. 13, no. 4, pp. 14–22, 2016. fict–based before–after study with use of comparison groups and the empirical Bayes method,” Transportation Research  U. Durrani, C. Lee, and D. Shah, “Predicting driver reaction Record, vol. 2659, no. 1, pp. 15–24, 2017. time and deceleration: comparison of perception-reaction thresholds and evidence accumulation framework,” Accident Analysis & Prevention, vol. 149, Article ID 105889, 2021.  K. El-Basyouny and T. Sayed, “A full Bayes multivariate in- tervention model with random parameters among matched pairs for before–after safety evaluation,” Accident Analysis & Prevention, vol. 43, no. 1, pp. 87–94, 2011.  B. N. Persaud, R. A. Retting, P. E. Garder, and D. Lord, “Safety efect of roundabout conversions in the United States: em- pirical bayes observational before-after study,” Transportation Research Record: Journal of the Transportation Research Board, vol. 1751, no. 1, pp. 1–8, 2001.  K. Huang, L. Wang, T. Tan, and S. Maybank, “A real-time object detecting and tracking system for outdoor night sur- veillance,” Pattern Recognition, vol. 41, no. 1, pp. 432–444,  S. Hasirlioglu and A. Riener, “Challenges in object detection under rainy weather conditions,” in Proceedings of the First International Conference on Intelligent Transport Systems, pp. 53–65, Guimarães, Portugal, November 2018.  E. Hauer, “Observational before/after studies in road safety,” Estimating the Efect of Highway and Trafc Engineering Measures on Road Safety, Pergamon Press, Kidlington, UK,  R. Elvik, H. Ulstein, K. Wifstad et al., “An Empirical Bayes before-after evaluation of road safety efects of a new mo- torway in Norway,” Accident Analysis & Prevention, vol. 108, pp. 285–296, 2017.  B. Persaud and C. Lyon, “Empirical Bayes before–after safety studies: lessons learned from two decades of experience and future directions,” Accident Analysis & Prevention, vol. 39, no. 3, pp. 546–555, 2007.  J. Park, M. Abdel-Aty, and J. Lee, “Use of empirical and full Bayes before–after approaches to estimate the safety efects of roadside barriers with diferent crash conditions,” Journal of Safety Research, vol. 58, pp. 31–40, 2016.
Journal of Advanced Transportation
Hindawi Publishing Corporation
A Before-and-After Study of a Collision Risk Detecting and Warning System on Local Roads
Journal of Advanced Transportation
, Volume 2023 –
Feb 28, 2023
Share Full Text for Free
Add to Folder
Web of Science